While machine learning (ML) has made enormous progress in recent years in hydrology, there are legitimate concerns that ML models may not be able to accurately represent future trends and unprecedented extremes, which are often underestimated. Recent physics-informed ML models, more specifically differentiable hydrologic models, seamlessly combine neural networks to physical prior knowledge like mass balances and conceptualized hydrologic fluxes to leverage the benefits of both worlds. In theory, such models may be better able to represent unseen scenarios. We put this theory to test by comparing traditional models, purely data-driven ML models, and differentiable models in terms of long-term future trends and extremes. Furthermore, we attempt to assess the impacts of both hydrologic models and meteorological forcing data on the simulated extremes. We show that both high-resolution forcing data and physical priors (the differentiable modeling framework) are beneficial to the simulation quality of the extreme events. For decadal-scale trend projection of flooding risks, physical priors are once again helpful in many situations, but the impact is reduced as hydrometeorological uncertainty appears to dominate. Both ML and physics-informed ML have advantages; this work highlights the different roles of forcings and model structure on extremes.